Biology subjects

Miller, L. M.

Publications and source records attributed to Miller, L. M..

11 recordsLinked to original sources

bioRxiv2026

Human-calibrated automated intelligibility evaluation of speech neuroprostheses

Objective: Brain-to-voice neuroprostheses can synthesize speech from the neural activity of people who have lost the ability to speak, enabling more natural and expressive communication. Assessing the intelligibility of synthesized speech, however, remains challenging because there is no ground-truth acoustic signal. The field instead relies on human transcription, which is costly and difficult to scale or compare across studies. Automatic speech recognition (ASR) offers a potential alternative, but state-of-the-art ASR models are optimized for transcription accuracy on typical fluent speech and may not reflect how humans perceive atypical synthesized speech. Approach: We collected human transcriptions for ~14.5 hours of synthesized speech from intracortical brain-to-voice decoders for one participant (yielding ~72 hours of transcribed speech). We used the human-transcribed synthesized speech to fine-tune Whisper-large-v3 with low-rank adaptation (LoRA) so that its word error rate (WER) approximates human WER. We then evaluated the model on held-out sentences. Main results: Fine-tuning increased the Pearson correlation between model WER and human WER from 0.62 to 0.76. The fine-tuned model's WER closely tracked human WER across evaluation sets, including speech synthesized by unseen brain-to-voice decoders. Close agreement with human WER was achieved with only 1.5 hours of synthesized speech. Significance: These results demonstrate that human-calibrated ASR can provide a scalable measure of brain-to-voice intelligibility that reflects human transcription more closely than zero-shot ASR does.

neuroscience↗

bioRxiv2026

Chirped Speech (Cheech) Enables Rapid Assessment of Multi-Level Auditory Evoked Potentials During Speech-in-Noise Recognition

Difficulties understanding speech in noise remain a common complaint even among listeners with normal hearing sensitivity, highlighting the need for objective, more effective measures of real-world listening. The goal of this study was to validate the use of a novel, chirped-speech (Cheech) stimulus - continuous, naturally-spoken speech fused with chirps designed to elicit robust auditory evoked potentials - to characterize relationships between speech recognition, listening effort, and auditory neural encoding. Twenty-five normal-hearing adults completed a sentence-recognition task using both original (unmodified) and Cheech-modified AzBio sentence lists in quiet, +3 dB, and -3 dB signal-to-noise ratio (SNR) conditions while neural responses from the brainstem through cortex were recorded simultaneously. Speech recognition remained near ceiling in quiet but declined with decreasing SNR for both original and Cheech stimuli. Compared with clean speech, Cheech-modified speech showed slightly poorer recognition performance as SNR decreased and somewhat higher perceived effort overall. Yet, Cheech was highly effective at evoking auditory responses from the brainstem (auditory brainstem response, ABR) through the cortex (including middle- and late-latency responses, MLR and LLR) even with <5 minutes listening time per condition. Neural responses showed reduced amplitudes and prolonged latencies as SNR decreased. In general, ABR latencies and wave I amplitudes were associated with speech-in-noise recognition performance, whereas cortical responses (MLR Na, Nb, and LLR P1) were associated with subjective workload. These findings show that Cheech-modified speech preserves intelligibility while yielding robust, multilevel neural recordings during sentence perception, offering a promising approach to examine hierarchical auditory processing under ecologically relevant speech-in-noise conditions.

neuroscience↗

bioRxiv2026

Brain2voice 2.0: High-performance voice synthesis brain-computer interface

Brain-computer interfaces (BCIs) offer a promising solution to speech loss due to neurological injury by decoding intended speech directly from brain activity. While recent BCIs have restored high-accuracy text-based communication, they fail to provide instantaneous voice output essential for the natural flow of conversation. Brain-to-voice BCIs address this gap by decoding voice directly from neural signals. However, even the state-of-the-art (SOTA) BCI-synthesized voice is not yet intelligible enough for real-world adoption. We introduce brain2voice 2.0, a new multimodal Transformer-based BCI decoder architecture capable of synthesizing highly intelligible voice from intracortical neural signals in real-time. Brain2voice 2.0 is trained on continuous and custom-tokenized acoustic targets and phoneme targets, leveraging their complementary speech information. We use self-supervised and adversarial training objectives that enhance acoustic feature quality and improve synthesis intelligibility. At each 10 ms timestep, the model causally outputs continuous and tokenized acoustic features for real-time voice synthesis as well as time-aligned phoneme predictions (raw phoneme error rate: 7%, comparable to the latest brain-to-text models). We evaluated this new approach on our prior intracortical brain-to-voice benchmark dataset (Wairagkar et al. 2025). Naive human listeners transcribed brain2voice 2.0 synthesized voice with a word error rate of 5.24%--an 8x improvement in intelligibility over previous SOTA results (43.75%). Brain2voice 2.0 demonstrates that highly intelligible real-time voice synthesis from neural signals is achievable, for the first time crossing the intelligibility threshold necessary for clinically viable brain-to-voice BCIs for people with paralysis.

neuroscience↗

bioRxiv2026

Comparison of a Novel Real-World Speech in Noise Auditory Attention Task to Standard Clinical Audiological Metrics

Pure tone audiometry (PTA) remains the clinical standard for evaluating hearing ability, yet individuals with similar audiometric profiles often exhibit substantial variability in their capacity to understand speech in everyday listening environments. Growing evidence suggests this variance is related to contributions from cognitive ability and auditory processing that standard threshold measures do not capture. To investigate how PTA, cognitive factors, and demographics such as age jointly predict real-world speech perception, 116 veteran adults 20-70 years old spanning a range of normal to moderate sensorineural hearing losses completed a spatial auditory attention task. Target color words were embedded within naturalistic short-story narratives presented under two conditions: a mono-talker speech-in-quiet (SIQ) condition and a dual-talker speech-in-noise (SIN) condition with a spatially separated competing narrative. Behavioral performance was quantified via color word hit accuracy, reaction time, and comprehension question accuracy. Participants also completed pure tone audiometry, the Montreal Cognitive Assessment (MoCA), and the Speech, Spatial and Qualities of Hearing Scale (SSQ12). Mixed-effects regression models were used to evaluate the contributions of PTA, age, cognitive ability, and self-reported hearing difficulty (SSQ12) to task performance across conditions. Results demonstrate a complex interplay between age, PTA, MoCA, and/or listening condition (SIQ vs. SIN) in predicting identification accuracy, reaction time, and comprehension. Age and condition significantly predicted hit accuracy and reaction time, with older participants showing improved accuracy in quiet but declining accuracy and slower responses in noise. PTA did not emerge as a significant main effect predictor but interacted with cognitive ability and condition to modulate performance, in some cases exhibiting a paradoxical inverse relationship with accuracy dependent on MoCA score. MoCA scores significantly predicted comprehension across conditions, and SIN hit accuracy was positively correlated with SSQ12 scores, validating the task against participants real-world listening experiences. These findings highlight the importance of incorporating cognitive screening and ecologically valid speech perception tasks into audiological assessment to better identify individuals at risk for functional hearing impairment in complex listening environments.

neuroscience↗

bioRxiv2026

Influenza A Virus Coinfection Alters Streptococcus pneumoniae Gene Expression during Upper Respiratory Tract Colonization

Streptococcus pneumoniae (Spn) asymptomatically colonizes the upper respiratory tract (URT), a niche from which it can transmit to another host or cause invasive disease in the same host. The in vivo transcriptional adaptations that Spn undergoes during nasopharyngeal colonization, particularly during influenza A virus (IAV) coinfection, are poorly understood. Here, we leveraged an established infant mouse model of colonization, shedding, and transmission to perform genome-wide transcriptomic profiling of Spn during mono- and during IAV coinfection. Compared with broth-grown controls, pneumococci isolated from the URT exhibited distinct transcriptional programs, with over 200 genes differentially expressed across time points. Genes involved in carbohydrate uptake and metabolism, glycan degradation, amino sugar and nucleotide sugar metabolism, and amino acid biosynthesis were consistently enriched during colonization, highlighting metabolic adaptation to the nasopharyngeal niche. In contrast, IAV coinfection induced a markedly distinct transcriptional signature, including upregulation of branched-chain amino acid biosynthesis, bacteriocin production, and phosphate acquisition systems. Notably, the pilus islet-1 locus was upregulated during Spn-IAV coinfection. Functional studies demonstrated that while the pilus was dispensable for colonization under mono- and coinfection conditions, it promoted high-shedding events and enhanced inflammatory responses during IAV coinfection. However, reduced inflammation and reduced high shedding events from pups inoculated with a pilus-deficient mutant did not alter transmission frequency in the infant mouse model. Collectively, our findings define the in vivo transcriptional landscape of Spn during URT colonization and reveal distinct bacterial adaptations during viral coinfection, providing insight into mechanisms that influence pneumococcal persistence, inflammation, and transmission.

microbiology↗

bioRxiv2024

Auditory brainstem responses to speech-in-noise reflect selective attention, comprehension, and subjective listening effort

The auditory brainstem plays a crucial role in speech-in-noise listening, refining numerous acoustic features such as pitch and spatial location under continual descending influence from cortex. However, the difficulty of characterizing brainstem activity during continuous speech listening has obscured its functional role in ecologically valid contexts--not only the effects of selective attention on neural responses, but also their impact on comprehension and listening effort. Here, we evaluate the role of the brainstem on speech-in-noise perception and selective attention using a continuous, speech-based stimulus with embedded chirps (Cheech) that rapidly and effectively evokes auditory brainstem responses (ABRs) while participants listen to short story narratives. The Cheech-modified stories were presented alone or in the presence of another spatially separated talker, and neural responses were measured throughout by EEG. Both word-level detection performance and narrative-level comprehension were evaluated, as well as subjective reports of listening effort. ABR wave V was modulated by the presence of a competing talker, selective attention for the target versus masker, and the talker sex. Additionally, faster wave V peak latencies and larger amplitudes were associated with identification of the target words embedded within the stories, while faster latencies additionally related to better comprehension question accuracy and lower subjective listening effort. Collectively, our results provide evidence for the influence of brainstem encoding processes on individual speech-listening behaviors, including the ability to selectively attend to and comprehend target speech in the presence of a competing talker. Significance statementThis study highlights the crucial role of the brainstem in speech-in-noise perception and higher-level language abilities like comprehension and subjective listening effort. Through the use of a novel speech stimulus blended with embedded chirps (Cheech) to evoke brainstem responses while listening to short story narratives, we show how selective attention influences subcortical processes. Faster and more robust subcortical processes, in turn, were associated with more successful speech-listening performance across multiple behavioral metrics. These findings advance our understanding of how the brainstem supports selective attention and speech perception in challenging listening environments.

neuroscience↗

bioRxiv2024

Individual Differences in Cognition and Perception Predict Neural Processing of Speech in Noise for Audiometrically Normal Listeners.

Individuals with normal hearing exhibit considerable variability in their capacity to understand speech in noisy environments. Previous research suggests the cause of this variance may be due to individual differences in cognition and auditory perception. To investigate the impact of cognitive and perceptual differences on speech comprehension, 25 adult human participants with normal hearing completed numerous cognitive and psychoacoustic tasks including the Flanker, Stroop, Trail Making, Reading Span, and temporal fine structure (TFS) tests. They also completed a continuous multi-talker spatial attention task while neural activity was recorded using electroencephalography (EEG). The auditory cortical N1 response was extracted as a measure of neural speech encoding during continuous speech listening using an engineered "chirped-speech" (Cheech) stimulus. We compared N1 component morphologies of target and masker speech stimuli to assess neural correlates of attentional gains while listening to concurrently played short story narratives. Performance on cognitive and psychoacoustic tasks were used to predict N1 component amplitude differences between attended and unattended speech using multiple regression. Results show inhibitory control and working memory abilities can predict N1 amplitude differences between the target and masker stories. Interestingly, none of the cognitive and psychoacoustic predictors correlated with behavioral speech-in-noise listening performance in the attention task, suggesting that neural measures may be more sensitive measures of cognitive and auditory processing as compared to behavioral measures alone. SIGNIFICANCE STATEMENTThese findings contribute to our understanding of how cognition affects the neural mechanisms of speech perception. Specifically, our results highlight the complex interplay between cognitive abilities and neural encoding of speech in challenging listening environments with multiple speakers. By incorporating these additional measures of cognition, we can achieve a more comprehensive understanding of an individuals speech perception abilities, even in individuals with normal hearing. This approach could lead to earlier detection of hearing issues and more personalized interventions, ultimately enhancing communication outcomes for those with hearing difficulty.

neuroscience↗

bioRxiv2024

Point mutation in a virus-like capsid drives symmetry reduction to form tetrahedral cages

Protein capsids are a widespread form of compartmentalisation in nature. Icosahedral symmetry is ubiquitous in capsids derived from spherical viruses, as this geometry maximises the internal volume that can be enclosed within. Despite the strong preference for icosahedral symmetry, we show that simple point mutations in a virus-like capsid can drive the assembly of novel symmetry-reduced structures. Starting with the encapsulin from Myxococcus xanthus, a 180-mer bacterial capsid that adopts the well-studied viral HK97 fold, we use mass photometry and native charge detection mass spectrometry to identify a triple histidine point mutant that forms smaller dimorphic assemblies. Using cryo-EM, we determine the structures of a precedented 60-mer icosahedral assembly and an unprecedented 36-mer tetrahedron that features significant geometric rearrangements around a novel interaction surface between capsid protomers. We subsequently find that the tetrahedral assembly can be generated by triple point mutation to various amino acids, and that even a single histidine point mutation is sufficient to form tetrahedra. These findings represent the first example of tetrahedral geometry across all characterised encapsulins, HK97-like capsids, or indeed any virus-derived capsids reported in the Protein Data Bank, revealing the surprising plasticity of capsid self-assembly that can be accessed through minimal changes in protein sequence. Significance statementViral capsids are cage-like protein assemblies that preferentially adopt icosahedral symmetry to maximise their internal volume for housing genetic material. This icosahedral preference extends to encapsulins, a widespread family of bacterial protein cages which evolved from viral capsids. Counter to this fundamental geometric preference, the formation of well-defined tetrahedral cages from a single amino acid substitution in an encapsulin reveals the surprising geometric flexibility of a common viral protein fold. These findings suggest that protein oligomerisation is far more permissive than intuitively expected, where serendipitous interactions between proteins arising from minimal mutations can cascade to form vast architectural changes. The ability to redesign protein architectures through simple mutations should enable biotechnological advances in vaccine development, drug delivery, and enzymatic biomanufacturing.

biochemistry↗

bioRxiv2022

Biophysical and Structural Characterization of a Viral Genome Packaging Motor

Like many dsDNA viruses, bacteriophage {lambda} replicates its genome as a concatemer consisting of multiple copies of covalently linked dsDNA genomes. To encapsidate a single genome within a nascent procapsid, {lambda} must: 1) find its own dsDNA amongst the multitude of host nucleic acids; 2) identify the genomic start site; 3) cut the DNA; 4) bring the excised DNA to a procapsid; 5) translocate DNA into the capsid; 6) cut DNA again at a packaging termination site, 7) disengage from the newly filled capsid; and 8) bring the remainder of the genomic concatemer to fill another empty procapsid. These disparate genome processing tasks are carried out by a single virus-encoded enzyme complex called terminase. While it has been shown that {lambda} terminase initially forms a tetrameric complex to cut DNA, it is not clear whether the same configuration translocates DNA. Here, we describe biophysical and initial structural characterization of a {lambda} terminase translocation complex. Analytical ultracentrifugation (AUC) and small angle X-ray scattering (SAXS) indicate that between 4 and 5 protomeric subunits assemble a cone-shaped terminase complex with a maximum dimension of [~]230 and radius of gyration of [~]72 [A]. Two-dimensional classification of cryoEM images of {lambda} terminase are consistent with these dimensions and show that particles assume a preferred orientation in ice. The orientations appear to be end-on, as terminase rings resemble a starfish with approximate pentameric symmetry. While [~]5-fold symmetry is apparent, one of the five "arms" appears partially displaced with weaker more diffuse density in some classes, suggesting flexibility and/or partial occupancy. Charge detection mass spectrometry (CDMS) is consistent with a pentameric complex, with evidence that one motor subunit is weakly bound. Kinetic analysis indicates that the complex hydrolyzes ATP at a rate comparable to the rates of other phage packaging motors. Together with previously published data, these results suggest that {lambda} terminase assembles conformationally and stoichiometrically distinct complexes to carry out different genome processing tasks. We propose a "symmetry resolution" pathway to explain how terminase transitions between these structurally and functionally distinct states.

biophysics↗

bioRxiv2022

Tween-20 induces the structural remodelling of single lipid vesicles

The interaction of Tween-20 with lipid membranes is crucial for a number of biotechnological applications including viral inactivation and membrane protein extraction, but the underlying mechanistic details have remained elusive. Evidence from ensemble assays supports a global model of Tween-20 induced membrane disruption that broadly encompasses association of the surfactant with the membrane surface, membrane fragmentation and the release of mixed micelles to solution, but whether this process involves intermediate and dynamic transitions between regimes is an open question. In search of the mechanistic origins of membrane disruption, increasing focus is put on identifying Tween-20 interactions with highly controllable model membranes. In light of this, and to unveil quantitative mechanistic details, we employed highly interdisciplinary biophysical approaches, including quartz-crystal microbalance with dissipation monitoring, steady-state and time-resolved fluorescence and FRET spectroscopy, dynamic light scattering, fluorescence correlation spectroscopy, wide-field single-vesicle imaging and scanning electron microscopy, to interrogate the interactions between Tween-20 and both freely-diffusing and surface-immobilized model-membrane vesicles. Using ultrasensitive sensing approaches, we discovered that Tween-20 leads to a stepwise and phase-dependent structural remodelling of sub-micron sized vesicles that includes permeabilization and swelling, even at detergent concentrations below the critical micellar concentration. These insights into the structural perturbation of lipid vesicles upon Tween-20 interaction highlight the impact on vesicle conformation prior to complete solubilization, and the tools presented may have general relevance for probing the interaction between lipid vesicles and a wide variety of disruptive agents.

biophysics↗

bioRxiv2021

Reconfigurable asymmetric protein assemblies through implicit negative design

Asymmetric multi-protein complexes that undergo subunit exchange play central roles in biology, but present a challenge for protein design. The individual components must contain interfaces enabling reversible addition to and dissociation from the complex, but be stable and well behaved in isolation. Here we employ a set of implicit negative design principles to generate beta sheet mediated heterodimers which enable the generation of a wide variety of structurally well defined asymmetric assemblies. Crystal structures of the heterodimers are very close to the design models, and unlike previously designed orthogonal heterodimer sets, the subunits are stable, folded and monomeric in isolation and rapidly assemble upon mixing. Rigid fusion of individual heterodimer halves to repeat proteins yields central assembly hubs that can bind two or three different proteins across different interfaces. We use these connectors to assemble linearly arranged hetero-oligomers with up to 6 unique components, branched hetero-oligomers, closed C4-symmetric two-component rings, and hetero-oligomers assembled on a cyclic homo-oligomeric central hub, and demonstrate such complexes can readily reconfigure through subunit exchange. Our approach provides a general route to designing asymmetric reconfigurable protein systems.

synthetic biology↗